Activation Steering: One Dial, Two Costs
Activation steering edits a model's internal activations to add or remove a concept at inference time. This explainer maps its one control knob, its two costs, and when it beats prompting.
Loading…
Topic
LLMs
658 posts about LLMs. Every summary links to the original.
- 658 posts
- 49 companies
- 5 explainers
See the idea
LLMs explainers
Antibodies as a Prediction Pipeline: Where AI Models Break Down
Antibody discovery is a three-stage prediction pipeline, and AI models that excel on familiar targets degrade sharply on unfamiliar ones. This explainer maps where that gap appears and what it costs.
Agent Memory: What Outlives the Run
Agent memory is the knowledge that survives a run, not the working state inside one. This explainer separates the two, then maps the write path, the read path, and the database limits that cap recall.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Posts
We got local models to triage the OpenClaw repo for FREE!*
OpenClaw receives hundreds of issues and pull requests daily, creating a triage problem for maintainers who need near-real-time filtering without consuming a paid closed-model quota. The system uses Gemma and Qwen local models inside a Pi-based agent harness: each item is normalized, enriched with GitHub context, classified into finite topic labels, and required to return structured output. For pull requests, the agent can inspect the repository through reposhell, a read-only command-restricted shell, while localpager, SQLite, a worker queue, and deterministic notification policy handle orchestration and Discord delivery. Tests described for gemma-4-26b-a4b and qwen3.6-35b-a3b report hundreds of tokens per second and good zero-shot classification accuracy without fine-tuning, while an LLM audit loop monitors errors; the authors present the pattern as agentic classification for high-throughput triage.
Dispatches from O'Reilly: From capabilities to responsibilities
High-stakes AI agents capable of mutating external state often face governance failures when relying on system prompts or manual Human-in-the-Loop approval queues that quickly degrade into alert fatigue. The Responsibility-Oriented Agent architecture addresses this operational bottleneck by shifting system design from open-ended capability framing to deterministic, contract-enforced responsibilities. Under this model, underlying orchestration frameworks like LangChain operate in User Space with their side-effecting tools removed, isolating the agent to epistemic reasoning. The agent expresses its intended action exclusively by emitting a structured policy proposal to a privileged Kernel Space runtime. The runtime deterministically evaluates the proposal against versioned YAML contracts registered in an agent registry, ensuring that only genuine policy exceptions are escalated to human supervisors.
Speculation Is All You Need
Speculative decoding is presented as the most important optimization for LLM inference, delivering speedups of 2x-3x rather than small percentage gains from kernel tuning. The technique accelerates the decode phase by using a lightweight draft model to generate candidate tokens that the target model processes in parallel, with lossless resampling. The authors released DFlash speculators for multiple Qwen models, achieving an additional 5-20% speedup over existing baselines, and worked with SGLang to integrate them. On a B200 node, the Qwen 3.5 122B-A10B model with speculation reaches over 1000 tok/s at concurrency 1, compared to 250 tok/s without speculation, demonstrating a 4x speedup. Training these speculators is simplified because the data-generating process is the target model itself, making data collection trivial, and the post envisions a flywheel where custom speculators and distillation enable continuous improvement in latency, cost, and quality for self-hosted inference.
Why everyone is talking about real-time analytics (a note from “the yellow company”)
Real-time analytics is presented as a foundational workload for customer-facing analytics, observability, fraud detection, operational applications, and AI agents rather than a niche use case. The discussion argues that performance claims require open, transparent, reproducible benchmarks that disclose methodology, datasets, queries, results, and operating conditions, including continuous ingestion, concurrency, memory pressure, and fresh data. A real-time database should support queryable data within milliseconds to a couple of seconds, incremental transformations, bounded tail latency as writes increase, configurable read-time versus write-time work, unified support for text, vector, JSON, and structured analytics, and efficiency as workload scales. The article connects renewed attention to agentic workloads, whose continuous, high-volume querying makes low latency, concurrency, and fresh context increasingly necessary, and calls for shared standards beyond a single latency number.
MosaicLeaks: Can your research agent keep a secret?
MosaicLeaks examines privacy leakage in deep-research agents that combine private enterprise documents with external web retrieval. Queries that appear harmless individually can reveal sensitive facts when an observer combines them over multiple hops, producing intent, answer, or full-information leakage. The benchmark contains 1,001 multi-hop chains interleaving local and web questions, and compares baseline behavior with task-only reinforcement learning and Privacy-Aware Deep Research (PA-DR). Task-only training raises strict chain success from 48.7% to 59.3% but increases answer/full-information leakage from 34.0% to 51.7%, whereas PA-DR reaches 58.7% success with 9.9% leakage by rewarding stage-specific decisions and penalizing query-level privacy risk. The evaluation is controlled, using synthetic enterprise documents, a fixed web corpus, three company contexts, and one multi-hop question-answering harness.
Data Engineering for AI: A Practical Guide for Data Professionals
Data engineering for AI extends traditional ETL by adding requirements for model-ready data, unstructured inputs, feature engineering, governance, and production monitoring. The guide addresses data professionals building or scaling AI infrastructure, covering ingestion, architecture, data quality, privacy compliance, generative AI, and career development. It emphasizes shared feature schemas, pipeline data contracts, lineage, statistical validation, drift detection, and PII stripping or anonymization as responsibilities shared across data engineering and data science. For generative AI systems, it describes preparing RAG pipelines by ingesting and chunking documents, creating vector embeddings, and indexing them for semantic retrieval, while evaluating vector databases for latency, scale, and integration. It concludes that reliable AI depends on fresh, accurate, compliant data and ongoing operational and architectural review.
The new bottleneck
AI coding tools have significantly lowered the cost of generating software, yet many engineering organizations fail to realize overall delivery speed improvements. Applying the Theory of Constraints reveals that eliminating code production as a bottleneck shifts inventory directly into surrounding legacy processes that remain unadjusted. New friction points consistently emerge across underspecified requirements, prolonged design handoff gates, senior engineer review capacity, and external sign-offs from legal or security. Organizations can address these blockers by interrogating legacy agile ceremonies checkpoint by checkpoint to determine if their original constraints still exist. Practical remedies include adopting real-time co-development between product and engineering, treating initial designs as fluid starting points, and restructuring processes around running rapid, high-volume experiments.
Beyond LoRA: Can you beat the most popular fine-tuning technique?
Parameter-efficient fine-tuning (PEFT) reduces the memory required to adapt open models, enabling fine-tuning with a fraction of the memory and, in some cases, quantized models. LoRA is dominant because it freezes base weights and trains a small set of added parameters, but its popularity may also reflect ecosystem momentum and uneven comparisons. The post presents shared-condition benchmarks for language-model math reasoning and image generation, using identical models, datasets, code, hardware, and measures including test performance, VRAM, forgetting or drift, runtime, and checkpoint size. The results show that LoRA performs well but is not consistently best: other techniques can outperform it on one or more axes, especially in the image-generation benchmark. The recommended choice is therefore to compare techniques against a use case’s needs and constraints, including quantization support, layer coverage, adapter merging, and downstream serving compatibility.
The open ecosystem around ClickHouse
ClickHouse’s open ecosystem spans clients, dashboards, ingestion pipelines, projects built on its engine, and agents that query it, with 2,600+ contributors, 240,000+ commits, and hundreds of millions of downloads cited. It describes open standards including OpenTelemetry for observability, Model Context Protocol (MCP) for AI tooling, and Iceberg and Delta Lake for lakehouse data, plus Iceberg REST catalogs, Apache Polaris, and Unity Catalog. The surrounding stack includes ClickStack, chDB, Langfuse, LibreChat, official language clients, a Kubernetes operator, Postgres bridges such as PeerDB and pg_clickhouse, and integrations with Airflow, dbt, Grafana, Fivetran, and Sigma. Specific updates include full Iceberg v2 support, Iceberg writes in 25.8, Paimon in 25.10, chDB v4’s DataStore API, and a Python client v1.0 with native async and 4x faster cold starts. The post frames this openness as supporting flexible tooling and community-led growth, reporting more than 200 integrations while noting that the ecosystem’s next direction remains unknown.
AI agents are a confused deputy with the keys to your kingdom
Attackers recently compromised over twenty thousand Instagram accounts by manipulating Meta's AI support assistant to rebind recovery email addresses without verifying account ownership. This incident illustrates the classic confused deputy security problem, where a privileged process is persuaded by an unprivileged user to perform unauthorized operations. Because large language model interfaces operate purely on natural language and cannot distinguish instructions from data, the model itself cannot serve as an authorization boundary. Securing AI agents requires verifying caller identity through external policy checks against authenticated sessions rather than relying on chat context or prompt engineering. Teams must enforce least privilege with short-lived scoped credentials, place irreversible actions behind hard policy gates or human approvals, and maintain audit trails of agent actions.
GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2 is presented as a flagship model for long-horizon tasks, extending GLM-5.1’s capability to a solid 1M-token context for sustained engineering work. It was trained extensively on 1M-context coding-agent scenarios involving large-scale implementation, automated research, performance optimization, and complex debugging, and adds effort-level controls to balance capability, latency, and computational cost. Its architecture uses IndexShare, sharing one lightweight indexer across every four transformer layers and reducing per-token FLOPs by 2.9× at 1M context; MTP changes for speculative decoding increase acceptance length by up to 20%. On reported long-horizon and standard coding benchmarks, GLM-5.2 is described as the highest-ranked open-source model: it trails Opus 4.8 by 1% on FrontierSWE, ranks second only to Opus 4.8 on PostTrainBench, and scores 81.0 on Terminal-Bench 2.1 versus GLM-5.1’s 63.5.
Introducing the Agentic CDP: A New Species of CDP for a New Era of Agents
Traditional customer data platforms (CDPs) were built for human-managed, batch-based campaigns, but the post argues that agentic buying requires millisecond speed, hyper-personalization, and richer context. It contrasts the familiar Golden Record with Golden Context, which combines customer data with current business goals and the history and outcomes of prior decisions. The proposed Agentic CDP uses “Infinity Campaigns,” always-on engagement loops that use LLMs and agents to adapt messaging, timing, and channels for individuals. It is also embedded in the data foundation, bringing customer, business, and decision context together under existing governance, and is designed for agents and humans from the outset. Databricks presents CustomerLake as an implementation of these principles for its platform.
Kimi K2.7 Code vs Claude Fable 5: Landing pages that cost 94% less
The experiment compares Kimi K2.7 Code and Claude Fable 5 across 12 landing-page prompts covering B2B SaaS, a rooftop speakeasy, developer tools, and other categories. Kimi cost about 94% less overall, averaging roughly 16 times cheaper than Fable, while producing similar-quality output, especially when given visual references. The team first established a prompt-only baseline, then used a custom design MCP server to provide screenshots, UI elements, and other references directly to multimodal Kimi; this improved hierarchy, typography, composition, readability, loading behavior, and image handling. GPT-5.5 scored screenshots and source code on a 0–100 rubric, and although Fable scored higher, the relatively small quality gap led the authors to consider Kimi’s lower cost a reasonable trade-off for iterative landing-page generation.
Clustering billions of products for agentic commerce with Catalog API
Shopify’s Catalog must reconcile billions of listings whose merchants use different product and variant structures, so AI shopping agents can identify the same real-world product. Product clustering groups related listings under a Universal Product Identifier, with precision prioritized over recall because wrong products can lead to incorrect purchases while missed variants remain findable. The intra-store system starts with rules-based singleton detection, then uses ANN and sparse average linkage for pre-chunking, and applies a two-stage LLM process that proposes clusters and critiques them using the core value proposition framework. Dynamic per-chunk strict JSON schemas enforce complete product coverage, support arbitrary chunking, and improved recall by 8% on the Toloka evaluation dataset in one schema change. The resulting pipeline is presented as a scalable intra-store foundation, while cross-store clustering into global UPIs is identified as the next step.
Building AI Agents for AR Glasses and XR Devices with NVIDIA XR AI
Building AI experiences for augmented reality and wearable devices requires bridging hardware with live media streams, multimodal models, enterprise tools, and runtime infrastructure. NVIDIA XR AI provides an open-source, modular framework connecting extended reality headsets and smart glasses to GPU-accelerated services across cloud, edge, and workstations. In this architecture, camera frames and microphone audio ingest into an XR Media Hub that routes data while keeping raw video pixels in shared memory to minimize overhead. The ecosystem uses NVIDIA Cosmos models for vision-language grounding, NVIDIA Nemotron models for reasoning and tool invocation, and the Model Context Protocol for enterprise integrations. Optional agent orchestration via NVIDIA NeMo Agent Toolkit and spatial streaming through NVIDIA CloudXR support complex workflows across healthcare and manufacturing.
NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance
NVIDIA submitted results across all MLPerf Training v6.0 benchmarks, setting performance records on workloads including the DeepSeek-V3 and GPT-OSS-20B Mixture of Experts models. To scale training up to 8,192 Blackwell GPUs, NVIDIA combined hardware cluster designs with scale-out networking via Spectrum-X Ethernet and Quantum InfiniBand. The engineering team eliminated CPU-GPU synchronization bottlenecks in token-dropless MoEs by implementing full-iteration CUDA graphs with synchronization-free operators and paged stashing. Further software optimizations included CuTe DSL kernel fusions, an MXFP8 attention block in the Transformer Engine, and transitioning router elementwise math to FP32. Across these benchmarks, the GB300 NVL72 platform achieved the fastest time to train at scale and the highest normalized per-accelerator performance.
Build On-Device AI Companions with the NVIDIA ACE Game Agent SDK and Unreal Engine 5 Plugins
Game developers integrating on-device artificial intelligence companions face challenges such as conversational latency, state synchronization, and loop prevention. NVIDIA announced the open-source ACE Game Agent SDK alongside a suite of Unreal Engine 5 plugins to streamline native, hardware-accelerated character pipelines. The SDK offers Agent, Chat, and RAG APIs, enabling characters to execute multi-step tool-assisted reasoning and query game databases. Unreal Engine 5 plugins introduce local runtime models including nemo-conformer-ctc-120m for speech recognition, Qwen 3.5 4B for language generation, and Chatterbox Turbo 350M for speech synthesis. Additional tooling expands to motion generation via Animotive Kimodo and rendering updates with the DLSS 4.5 Unreal Engine plugin.
How to Optimize Transformer-Based Models for Low-Precision Training
Accelerating transformer training requires understanding how low-precision formats such as FP8 and NVFP4 impact specific general matrix multiplication (GEMM) workloads. Transformer configurations do not explicitly reveal these shapes, making microbenchmarks necessary before committing to full training runs. NVIDIA Transformer Engine enables quantization and kernel dispatch across precisions, which can be evaluated in realistic autocast mode or kernel-only prequantized mode. Profiling the ESM2-15B model on NVIDIA B300 GPUs demonstrated that NVFP4 achieved a 1.79x blended forward propagation speedup over MXFP8 and up to 4.01x over BF16 in prequantized execution. Although large GEMM dimensions successfully overcome quantization overheads, dynamic scaling, Hadamard transforms, and kernel selection asymmetries moderate real-world gains.
Salesforce signs definitive agreement to acquire Fin
Salesforce has signed a definitive agreement to acquire Fin for approximately $3.6 billion, with closing expected in the fourth quarter of Salesforce’s fiscal year 2027. Fin says it began as Intercom 15 years ago and, nearly four years ago, shifted toward modern LLMs to create and define the Customer Agents category. The companies plan to join forces, with Salesforce’s resources intended to help bring Fin to more consumers, while Fin says little will practically change for customers: the author will remain CEO and Des Traynor will continue running R&D. The announcement also points to recent work including the Apex model and the Operator internal agent, and describes the deal as a pivotal moment for Fin’s employees, founders, and leadership team.
How Wayfair cut ML model costs by 90% (twice!) with Cursor
Wayfair’s Applied Research team used Cursor to automate and parallelize experiments for a product-tag validation model that checked attributes against images, descriptions, and reviews. The model was accurate but too expensive to run across the catalog. In December 2025, five researchers standardized the dataset and evaluation benchmark, then built and tested 110 distinct variants in four days; the winning architecture reduced inference costs by 94%, improved precision, and became the production baseline. In March 2026, the team benchmarked against that baseline, ran 140+ new experiments, and added genetic-algorithm searches, achieving another 90% cost reduction. The workflow used 20+ parallel agents, desktop and CLI surfaces, and cloud agents, leaving researchers to formulate hypotheses, interpret results, and choose candidates while Cursor implemented, ran, and measured variants.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.